Third SRNWP Workshop on Short-range EPS

10-11 December 2007, Rome (Italy)



Report of the final discussion

Tuesday, 15h45-17h00

Chair: Peter Houtekamer


Peter Houtekamer, Massimo Bonavita and Jean Quiby have prepared the following questions for the workshop final discussion.



I. The different strategies for the initialization of the LAMs:

downscaling from global, singular vectors, breeding, EnKF data assimilation


Downscaling from global

Although we have the advantage of a better representation of the soil variables, the majority of the participants agrees that with this technique not much value is added with respect to the global ensemble that is used to pilot the LAM ensemble. This could be different for mountainous areas, where the higher horizontal model resolution of the LAMs permits a steeper model orography, which would in particular improve the prediction of strong advective precipitation.

Otherwise, no supplementary dynamical (e.g. perturbations) or meteorological (e.g. observations) information is given to the LAMs.



As already assessed at the Second Workshop, it has been repeated that breeding is a method of limited sophistication, because the perturbations cannot be made orthogonal. A great advantage of the breeding technique is its simplicity when compared to the singular vector technique.


Singular vectors

The big advantage of this method, with respect to the breeding method, is that it defines orthogonal perturbations. It is generally admitted today that the SV method is an appropriate method for global models. In fact, after two days of integration, it would appear that perturbations project on fairly similar error structures for the breeding, singular vector and EnKF-based ensemble initialization methods.


The precise way of initializing an ensemble is likely more important for high-resolution limited area applications. High resolution LAMs are used for short range, even for very short range forecasting. For these short ranges, you would like to have the perturbations fully developed from the beginning of the integration. But when an optimization time of only a few hours is used, the eigenvalue spectrum of the SV perturbations is very flat.

Another weakness of the SV method for the LAMs is that the highly non-linear diabatic processes are not (yet?) considered in the computation of the perturbations, although diabatic processes can be locally very important.

It has been strongly advocated by a participant that we have to stick to the SV method because its rapidly growing dynamical perturbations capture some of the error growth that is actually due to model error. At this time, a more appropriate simulation of model errors remains to be developed. (Whether we will one day be able to simulate model errors is not certain).


EnKF data assimilation

This method was considered important. It is a pity that it has not received in Europe the importance that it deserves. However the situation is presently changing: ECMWF is considering it very seriously as possible alternative and the NWP Consortium COSMO (among other countries Germany and Italy) has decided to start its development. For an EnKF that is operating on a local domain, it is not clear how to deal with boundary perturbations. No experiments have as yet been performed with coupled global and regional EnKF systems.



II. How many members do we need for LAM EPS?


Olivier Talagrand has claimed that probabilistic scores saturate in the range of 20-50 members and consequently it is difficult to justify having more than 50 members. This statement has not been accepted unanimously, because it was felt that low-probability warnings against extreme events are important for users with a low cost-loss ratio. Such users might act on a warning by 10 out of 500 members but not on a warning by just 1 out of 50 members.

Note that the European Project GLAMEPS foresees the use of a very large ensemble, as it will also probably be the case with TIGGE-LAM.

It has been said that we should also look at the practical work needed for the development of an EPS. An EPS with, say, 200 members would be practically impossible to validate. Alternatively, one could use available computer resources to improve the realism of individual members by, for instance, increasing their horizontal or vertical resolution.


III. How to best account for model errors?

Should we add random errors, use different parametrizations, implement stochastic physics or stochastic backscattering of kinetic energy?


Concerning the high resolution LAMs, one of the strongest statements made in the discussion of this topic has been the following: Whatever you do inside the integration domain, this will less influence the solution than a modification in the boundary conditions.


The pragmatic multi-model approach comprehensive samples a wide variety of model errors. However, it is hard to maintain an ensemble of significantly different models of equal quality. It might be preferable to have some sort of super-parameterization of model error. It is not clear, however, how to proceed with this. The main reason must have been that until today, there have not been enough studies, we have not accumulated enough experience to assess the respective merits and weaknesses of theses different techniques.

A major difficulty for answering the above question is that the cause of model errors is in many cases not clearly identifiable. Examples: when several causes mix, when two parametrizations are incompatible, when the coupling between dynamics and physics is inconsistent. Once a source of errors has been identified, it will often not be simulated but instead be reduced by subsequent research and development .


IV. Postprocessing of the EPS forecasts:

a). Is a recommended way of doing it?

b). Can the calibration of a single model EPS be a substitute for a calibrated multi-model EPS?


For this topic, the discussion has been lean.

For the question a), nobody could make a recommendation.

Concerning question b), the lack of statements of the participants seemed to be caused by the fact that, from the presentations given during the workshop, contradictory results have been presented.

Nevertheless, it came out of the discussion that multi-model ensembles are not a guarantee for good spread and that reliable single model ensembles are not necessarily less skilful than multi-model ensembles. Two very important statements!



For the minutes:

Jean Quiby

With thanks to Pieter Houtekamer for the reviewing and the enhancement of this report